关于「术语图:锁表、同义漂移与跨卷一致性审计」,《多窗口写书防漂移》给出的定量结论是什么?
结论:术语漂移有字面变体、同义替换与对译漂移三种来源,字符串尺子只能抓第一种,第二种靠锁表里登记的别名,第三种靠中英配对构成的术语图。编辑距离在短中文术语上一字千钧,两字术语差一个字相似度就只剩 0.500;字二元组要补边界,英文要先比紧凑键。构造样本上,两把尺子在召回率接近满分时精确率只有六成左右,所以它们只负责产生复核名单,自动判定一律来自锁表。跨卷术语图只沿硬边求分量,软边从不自动合并;五条审计规则把碰撞与对译漂移放在最前,定义冲突交还定义权所在的那一卷。锁表每次改动的波及面,按「用了且版本钉早于修改」精确计算。下一章转向与术语同样会在多窗口之间漂移的另一样东西:数字。
数据口径:书中未标出处的错误率都是假设参数或合成数据上的测量值;本书不是法律意见,也不推荐任何写作工具或模型。 第八章让四个窗口并发写一本九章合成书,注入五类共 119 个错误:四道闸门拦住 0.723,漏网的 33 个里有 28 个属于机器在结构上看不见的手法,扣除后检出率为 0.945;干净书上术语检测首轮误报 1672 处,人审 48 个词后归零;一份严肃的注入审计每类至少要 75 个错误。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 与结论吻合的打印值: 0.500
| DP table for pair Z5 (rows: prefixes of the 4-character term, columns: of the 5-character term) | |
| row 0 | 0 1 2 3 4 5 |
| row 1 | 1 0 1 2 3 4 |
| row 2 | 2 1 0 1 2 3 |
| row 3 | 3 2 1 0 1 2 |
| row 4 | 4 3 2 1 1 1 |
| pair (first / second) len d sim | |
| Z1 (Chinese pair) | 5 1 0.800 |
| Z2 (Chinese pair) | 5 2 0.600 |
| Z3 (Chinese pair) | 2 1 0.500 |
| Z4 (Chinese pair) | 3 1 0.667 |
| Z5 (Chinese pair) | 5 1 0.800 |
| Z6 (Chinese pair) | 4 1 0.750 |
| Z7 (Chinese pair) | 3 2 0.333 |
| Z8 (Chinese pair) | 4 1 0.750 |
| E1 exclusive-create lock / exclusive create lock | 21 1 0.952 |
| E2 exclusive-create lock / exclusive lock | 21 7 0.667 |
| E3 lease / lease term | 10 5 0.500 |
| E4 heartbeat / heart-beat | 10 1 0.900 |
| E5 term lock table / term table | 15 5 0.667 |
| E6 calculation box / hypothesis box | 15 10 0.333 |
| E7 citation hallucination / citation fabrication | 22 6 0.727 |
| E8 stale-lock reclamation / stale lock recovery | 22 9 0.591 |
| one edit on a 2-character term: sim | 0.500; on a 21-character term: sim = 0.952 |
| swap typo 'term-lcok' vs 'term-lock': Levenshtein 2, OSA 1 |
查看完整验算程序
# Chapter 3. p30: Levenshtein distance on invented term pairs; one edit weighs more on a short Chinese term.
# Wagner-Fischer dynamic programming: insert, delete, substitute each cost 1. Normalised similarity
# sim = 1 - d / max(len a, len b). The OSA variant also counts one swap of adjacent characters as 1.
# All terms are invented for this book's fictitious series; labels are not used here (see p32).
def lev(a, b):
prev = list(range(len(b) + 1))
for i, ca in enumerate(a, 1):
cur = [i]
for j, cb in enumerate(b, 1):
cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ca != cb)))
prev = cur
return prev[-1]
def osa(a, b):
d = [[0] * (len(b) + 1) for _ in range(len(a) + 1)]
for i in range(len(a) + 1):
d[i][0] = i
for j in range(len(b) + 1):
d[0][j] = j
for i in range(1, len(a) + 1):
for j in range(1, len(b) + 1):
c = a[i - 1] != b[j - 1]
d[i][j] = min(d[i - 1][j] + 1, d[i][j - 1] + 1, d[i - 1][j - 1] + c)
if i > 1 and j > 1 and a[i - 1] == b[j - 2] and a[i - 2] == b[j - 1]:
d[i][j] = min(d[i][j], d[i - 2][j - 2] + 1)
return d[-1][-1]
def table(a, b):
rows = [[0] * (len(b) + 1) for _ in range(len(a) + 1)]
for i in range(len(a) + 1):
for j in range(len(b) + 1):
rows[i][j] = lev(a[:i], b[:j])
return rows
PAIRS = [
("独占新建锁", "独占创建锁"), ("独占新建锁", "排他新建锁"), ("租约", "租期"), ("心跳", "心跳包"),
("术语锁表", "术语锁定表"), ("术语锁表", "术语表"), ("计算框", "假说框"), ("引文幻觉", "引用幻觉"),
("exclusive-create lock", "exclusive create lock"), ("exclusive-create lock", "exclusive lock"),
("lease", "lease term"), ("heartbeat", "heart-beat"), ("term lock table", "term table"),
("calculation box", "hypothesis box"), ("citation hallucination", "citation fabrication"),
("stale-lock reclamation", "stale lock recovery"),
]
if __name__ == "__main__":
a, b = "术语锁表", "术语锁定表"
print("DP table for pair Z5 (rows: prefixes of the 4-character term, columns: of the 5-character term)")
for i, row in enumerate(table(a, b)):
print(" row {}: {}".format(i, " ".join(str(x) for x in row)))
print(" pair (first / second) len d sim")
for x, y in PAIRS:
d = lev(x, y)
n = max(len(x), len(y))
k = PAIRS.index((x, y))
lab = "E{} {} / {}".format(k - 7, x, y) if x.isascii() else "Z{} (Chinese pair)".format(k + 1)
print(" {:<53} {:>3} {:>2} {:.3f}".format(lab, n, d, 1 - d / n))
print("one edit on a 2-character term: sim = {:.3f}; on a 21-character term: sim = {:.3f}".format(
1 - 1 / 2, 1 - 1 / 21))
print("swap typo 'term-lcok' vs 'term-lock': Levenshtein {}, OSA {}".format(
lev("term-lcok", "term-lock"), osa("term-lcok", "term-lock")))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《多窗口并行写书防漂移控制、术语跨卷一致性图审计与零幻觉文献核验器》第 3 章「术语图:锁表、同义漂移与跨卷一致性审计」(版本 0.1.0). 5D Library. DOI 待发布
DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-152 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026152,
author = {Fan, Yafu},
title = {Drift Control for Parallel Multi-Window Book Writing, Cross-Volume Terminology Graph Audits and Zero-Hallucination Citation Verifiers},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}